Power grid fault prediction and risk assessment method based on deep learning

Abstract

The invention discloses a power grid fault prediction and risk assessment method based on deep learning, and belongs to the technical field of power grid faults, and the method comprises the steps: determining deviation classification and a data scheduling range corresponding to the deviation classification according to power grid region information; setting a mother prediction model for each deviation classification; performing real-time optimization on the mother prediction model based on the fault detail list; segmenting the power grid region according to the data scheduling range to form a prediction region and a region data range; setting a sub-prediction model for the prediction area according to the mother prediction model; after the parent prediction model is optimized, updating the child prediction model corresponding to the parent prediction model; monitoring data of a power grid area are obtained in real time, the monitoring data are classified according to the area data range of a prediction area, area monitoring data of the prediction area are formed, the area monitoring data are analyzed through the sub-prediction model, and a fault prediction result is obtained; and performing risk estimation according to the fault prediction result to obtain a risk prediction result.

Description

Technical Field

[0001] The present invention belongs to the technical field of power grid failure, and specifically is a power grid failure prediction and risk assessment method based on deep learning. Background Art

[0002] In the current power system operation and maintenance management, fault prediction and risk assessment are key links to ensure the stable operation of the power grid. Traditionally, the prediction of power grid faults mainly relies on manual experience analysis and regular inspections. This method is not only inefficient, but also difficult to fully cover all potential fault risks. With the rapid development of big data, artificial intelligence and machine learning technologies, it has become possible to build a fault prediction model to achieve intelligent prediction of power grid faults.

[0003] Although some machine learning-based fault prediction models have been applied to the power grid, most of these models use a unified global model to process the data of the entire power grid. However, the power grid has a complex network structure and diverse operating conditions, and the fault types and characteristics of different regions or subnets often vary significantly. Therefore, the prediction accuracy and generalization ability of the global model are limited when processing such heterogeneous data.

[0004] In addition, existing fault prediction models often lack dynamic self-optimization mechanisms. Once the model is established, its parameters and identification features are relatively fixed, making it difficult to adapt to the changing operation status of the power grid and the emergence of new fault modes. This results in a significant drop in the prediction performance of the model when faced with new fault types or changes in identification features.

[0005] Based on this, in order to solve the above problems, the present invention provides a power grid fault prediction and risk assessment method based on deep learning. Summary of the invention

[0006] In order to solve the problems existing in the above-mentioned solutions, the present invention provides a power grid fault prediction and risk assessment method based on deep learning.

[0007] The purpose of the present invention can be achieved through the following technical solutions:

[0008] A power grid fault prediction and risk assessment method based on deep learning, the method comprising:

[0009] Step 1: Obtain grid area information, and determine deviation classification and data dispatch range corresponding to the deviation classification according to the grid area information;

[0010] Furthermore, the method for partitioning the power grid area according to the power grid area information includes:

[0011] Obtaining historical prediction records of power grid faults by each application model, wherein the application model is a power grid fault prediction model of historical application; identifying power grid faults of the power grid according to the historical prediction records, and marking the power grid faults as potential faults;

[0012] Determine a deviation classification according to the historical prediction record, wherein the deviation classification is composed of potential faults, and the potential faults in different deviation classifications are different;

[0013] A fault detail table is set according to the deviation classification, wherein the fault detail table is used to count potential faults and fault identification features corresponding to the potential faults;

[0014] Mark corresponding potential faults for each grid device in the grid area according to the grid area information and the fault details table; determine the collection scope of the grid device marking the potential fault according to the fault identification characteristics of the potential fault;

[0015] The collection range of the power grid equipment is integrated according to the deviation classification to obtain the data scheduling range corresponding to the deviation classification.

[0016] Furthermore, the method for determining the deviation classification based on the historical prediction records includes:

[0017] Calculate the prediction accuracy of the application model for potential faults through historical prediction records; identify the highest prediction accuracy of the potential faults and mark it as the benchmark accuracy; calculate the standard interval of the potential faults according to the benchmark accuracy and a preset allowable deviation value; classify the potential faults according to the standard interval to obtain an initial classification corresponding to the application model, the initial classification includes potential faults whose prediction accuracy of the application model is within the standard interval; remove duplicates from the initial classification;

[0018] Setting a combination standard, wherein the combination standard includes all potential faults; combining according to the combination standard and the initial classification to form a candidate combination that meets the combination standard;

[0019] The candidate combinations are screened to determine a target combination; and the target combination is simulated and adjusted to obtain a deviation classification.

[0020] Step 2: Set up a power grid fault prediction model for each deviation classification, marked as the parent prediction model; obtain a fault detail list, and optimize the parent prediction model in real time based on the fault detail list; divide the power grid area according to the data scheduling range of the deviation classification to form several prediction areas corresponding to the deviation classification, identify the data scheduling range corresponding to the prediction area, and mark it as the regional data range.

[0021] Furthermore, the method for segmenting the power grid area according to the data dispatching range of the deviation classification includes:

[0022] Step SA1: Divide the power grid area into several unit areas, identify the collection range of each power grid device in each unit area, and integrate them into a unit range; estimate the data analysis amount of the unit area according to the unit range; set a limit on the data amount for each deviation classification;

[0023] Step SA2: Select any unit area as the initial area; when there is no initial area, end the analysis;

[0024] Step SA3: Mark any unit area adjacent to the initial area as a candidate area;

[0025] When there is no area to be selected, the initial area is marked as the prediction area, and the process returns to step SA2;

[0026] Step SA4: Add the data analysis amount of the initial area and the data analysis amount of the selected area to obtain the amount of data to be merged, and compare the amount of data to be merged with the limited amount of data;

[0027] Step SA5: when the amount of data to be merged is greater than the limited amount of data, cancel the mark of the selected area and return to step SA3;

[0028] When the amount of data to be merged is not greater than the limited data amount, the initial region is merged with the selected region to form a new initial region, the amount of data to be merged is marked as the data analysis amount of the new initial region, and the process returns to step SA3.

[0029] Furthermore, the method for real-time optimization of the parent prediction model based on the fault list includes:

[0030] The fault detail table is updated in real time, and newly added faults are identified in real time in the fault detail table; the occurrence probability and fault impact value of each newly added fault are determined in real time, and the value range of the fault impact value is [0, 100];

[0031] Assign the newly added faults to corresponding deviation categories, mark the newly added faults in the deviation categories as i, i=1, 2, ..., n, and n is the number of newly added faults; mark the occurrence probability and fault impact value as λi and GZi respectively;

[0032] According to the formula Calculate the optimized value in real time; where: YP is the optimized value;

[0033] Determine whether to optimize the parent prediction model corresponding to the deviation classification according to the optimization value;

[0034] When it is determined that the parent prediction model is to be optimized, optimizing the parent prediction model according to the newly added fault;

[0035] When it is determined that the parent prediction model is not to be optimized, no optimization operation is performed.

[0036] Step 3: Setting a corresponding sub-prediction model for the prediction area according to the parent prediction model classified by the deviation; and after the parent prediction model is optimized, updating the sub-prediction model corresponding to the parent prediction model;

[0037] Step 4: Acquire monitoring data of the power grid area in real time, classify the monitoring data according to the regional data range of the prediction area, form monitoring data of the prediction area, mark it as regional monitoring data, analyze the regional monitoring data through the sub-prediction model, and obtain fault prediction results.

[0038] Step 5: Perform risk estimation based on the fault prediction result to obtain a risk prediction result.

[0039] Furthermore, the method for performing risk estimation based on the fault prediction result includes:

[0040] Identify the target fault according to the fault prediction result, and generate a fault probability curve of the target fault, wherein the horizontal axis of the fault probability curve is time and the vertical axis is the fault occurrence rate;

[0041] Fitting the fault probability curve to obtain a fault probability function, marking the fault probability function as H(t), where t is time;

[0042] Determine the fault impact value of the target fault; calculate the risk value according to the risk assessment formula; and form a risk prediction result based on the risk value.

[0043] Furthermore, the risk assessment formula is:

[0044] Where: FB is the risk value; GZ′ is the fault impact value; T is the interval time.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] By constructing a hierarchical architecture of parent prediction models and child prediction models, accurate prediction and efficient risk assessment of power grid faults are achieved. The system not only significantly improves the accuracy and timeliness of fault prediction, but also provides strong decision-making support for the operation and maintenance management of the power grid, and has important practical application value. By updating the fault table in real time, new faults and different identification features are dynamically determined according to the fault identification type and identification features of the parent prediction model, so that the system can adapt to the changing state of the power grid. The parent prediction model is iteratively optimized by optimizing data, and the generalization ability and prediction accuracy of the model are continuously improved to ensure that the system can maintain stable prediction performance in various complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0048] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0049] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] like Figure 1 As shown, a power grid fault prediction and risk assessment method based on deep learning, the method includes:

[0051] Step 1: Obtain the grid area information, and determine the deviation classification and the data dispatch range corresponding to the deviation classification according to the grid area information;.

[0052] The grid area information refers to various information corresponding to the grid area, such as various grid equipment information, location, natural environment and other related data.

[0053] The deviation classification and data scheduling range are determined by partitioning according to the limitations of the intelligent model based on deep learning. The power grid has a complex network structure and diverse operating conditions. The fault types and characteristics in different regions or subnets often differ significantly, resulting in a decrease in the fault prediction accuracy of the global model in certain situations. Therefore, it is necessary to partition according to actual conditions to achieve targeted fault prediction.

[0054] In one embodiment, the method for determining deviation classification and data scheduling range is: partitioning based on existing technology.

[0055] In one embodiment, the method for determining the deviation classification and data scheduling range is as follows:

[0056] The existing intelligent model for power grid fault prediction is marked as an application model; the historical prediction records of the application model for power grid faults are obtained, and the historical prediction records include prediction results, actual results, fault types, faulty equipment and other related data; various power grid faults in the current power grid are determined according to each historical prediction record and marked as potential faults; different fault identification features of the same fault result are regarded as different potential faults, such as the fault result of insulator self-explosion, which has multiple situations leading to self-explosion, and different situations correspond to different potential faults.

[0057] Determine each deviation category based on each historical prediction record. The deviation category consists of various potential faults. That is, the power grid faults belonging to the same deviation category can be predicted through the same application model to ensure the accuracy of the prediction. The power grid faults in different deviation categories are predicted through the same application model, which will lead to a decrease in the accuracy of the prediction of power grid faults in other deviation categories.

[0058] A fault detail table is set according to each deviation classification. The fault detail table is used to count various potential faults and the fault identification features corresponding to the potential faults;

[0059] According to the grid area information and the fault details table, the corresponding potential faults are marked for each grid device in the grid area, that is, it is determined whether the potential fault occurs in the grid device. If the potential fault occurs, a comparison is made; according to the fault identification characteristics corresponding to the potential fault, the monitoring data required for the grid device to predict the potential fault is determined, and the collection scope of the potential fault for the grid device is formed, and data is subsequently collected according to the collection scope;

[0060] The collection scope of each power grid device is integrated according to each deviation classification to obtain the data dispatching scope corresponding to each deviation classification, that is, the collection scope of each power grid device for each potential fault in the deviation classification is integrated.

[0061] In one embodiment, each deviation classification is determined according to each historical prediction record. This can be determined based on existing technologies. Because in the use of each application model, it is generally summarized under which circumstances the failure prediction of potential failures is not ideal. Therefore, a direct corresponding result summary can be made to form each deviation classification.

[0062] In one embodiment, the method for determining each deviation classification according to each historical prediction record includes:

[0063] The prediction accuracy of various application models for each potential fault is statistically analyzed through historical prediction records. If the corresponding potential fault cannot be predicted, no statistics will be performed.

[0064] Identify the highest prediction accuracy of each potential fault and mark it as the benchmark accuracy. Predict the allowable deviation value, that is, the value below the benchmark accuracy that can be considered a targeted fault prediction, which is set by the platform. Form a standard range for each potential fault based on the benchmark accuracy and allowable deviation value of each potential fault. Screen and classify the potential faults of each application model according to each standard range to obtain the initial classification corresponding to each application model, that is, the initial classification is the potential faults whose prediction accuracy of the application model is within the corresponding standard range, indicating that the application model is only suitable for fault prediction of each potential fault within the initial classification. De-duplicate each initial classification.

[0065] A combination standard is set, and the combination standard includes all potential faults; each initial classification is combined according to the combination standard to form each candidate combination that meets the combination standard, that is, each candidate combination includes all potential faults.

[0066] Screen each candidate combination and determine the target combination; that is, screen based on the number of initial classifications in the candidate combination. The smaller the number, the higher the priority, because the smaller the number, the fewer intelligent models will be established subsequently; if the number is the same, determine the priority based on other factors such as the estimated difficulty of implementation and establishment; you can also screen based on existing technologies.

[0067] Identify each initial classification in the target combination and mark it as the target classification;

[0068] Simulate and adjust each target classification in the target combination to obtain each deviation classification; that is, use the existing fault prediction technology to simulate it, determine whether classification optimization can be performed again, and mark the remaining target classification after optimization as deviation classification.

[0069] Step 2: Based on deep learning, a power grid fault prediction model is set for each deviation classification, marked as the parent prediction model; the power grid fault prediction model is used to analyze the monitoring data corresponding to each potential fault, and estimate whether the potential fault will occur. If it will not occur within the preset time period, the fault prediction result is normal operation. If the potential fault will occur within the preset time period, the fault prediction result is the probability of fault occurrence at each time in the subsequent predicted time.

[0070] Obtain a detailed fault list and optimize the parent prediction model in real time based on the detailed fault list; divide the power grid area according to the data dispatch range of each deviation classification to form each prediction area, identify the data dispatch range corresponding to the prediction area, and mark it as the regional data range.

[0071] In one embodiment, each parent prediction model is established based on the existing technology.

[0072] For example, data is collected in real time from sensors and measuring devices deployed on predetermined nodes of the power grid and transmission lines, including parameters such as current, voltage, temperature, and meteorology. These data will be used to train the deep learning model. The collected data is cleaned, outliers and missing values ​​are removed, and normalized to ensure the accuracy and consistency of the data. In addition, data enhancement technology can be used to increase the sample size of different fault conditions in the data set by simulating the upcoming data changes of the power grid under different fault conditions. According to the needs of the project and the characteristics of the data, a suitable deep learning framework such as TensorFlow or PyTorch is selected. A deep learning network model such as convolutional neural network (CNN), recurrent neural network (RNN) or long short-term memory network (LSTM) is designed. These network structures can extract features from data and make predictions. Use technologies such as autoencoders to extract multi-dimensional features of power grid operation from preprocessed data, including current intensity, voltage level change patterns, and periodic changes in load flow. Use the extracted features as input to train the deep learning model. Optimize the prediction performance of the model by adjusting the parameters and structure of the model. During the training process, techniques such as cross-validation can be used to evaluate the generalization ability of the model.

[0073] In one embodiment, the power grid area is divided according to the data scheduling range of each deviation classification, which is determined according to the data processing amount of the power grid fault prediction model, that is, the power grid area is divided according to the comprehensive data processing efficiency, prediction time requirements, etc. to form various prediction areas; that is, the prediction areas corresponding to different deviation classifications may not be the same; specifically, the division can be based on existing technologies.

[0074] In one embodiment, the method for segmenting the power grid area according to the data dispatch range of each deviation classification includes:

[0075] Step SA1: Divide the power grid area into several unit areas according to the specified minimum area. The minimum area can be a specified area such as a substation area, an administrative area, etc., but the premise is that the data between the unit areas are separated. For example, the data required to analyze potential problem A are all in the same unit area and will not be separated in different unit areas;

[0076] Identify the collection range of each power grid device in each unit area, integrate them into a unit range, and estimate the data analysis volume of the unit area based on the unit range. The data analysis volume is the amount of data that needs to be analyzed. It can also be expressed in terms of data such as the required processing time that can represent the data processing situation.

[0077] A data limit for the power grid fault prediction model is set for each deviation category, that is, the upper limit of the data analysis volume in the responsible area, which is mainly set by the platform according to the user's prediction efficiency requirements for the deviation category;

[0078] Step SA2: Select any unit area as the initial area; when there is no initial area, end the analysis;

[0079] Step SA3: Mark any unit area adjacent to the initial area as a candidate area;

[0080] When there is no area to be selected, the initial area is marked as the prediction area, and the process returns to step SA2;

[0081] Step SA4: Add the data analysis amount of the initial area and the data analysis amount of the selected area to obtain the amount of data to be merged, and compare the amount of data to be merged with the limited amount of data;

[0082] Step SA5: when the amount of data to be merged is greater than the limited amount of data, no merging is performed and the mark of the selected area is cancelled, that is, the unit area will not be used as a selected area of ​​the initial area in the future, and the process returns to step SA3;

[0083] When the amount of data to be merged is not greater than the limited data amount, the initial region is merged with the selected region to form a new initial region, the amount of data to be merged is marked as the data analysis amount of the initial region, and the process returns to step SA3.

[0084] In one embodiment, a method for real-time optimization of a parent prediction model based on a fault detail list includes:

[0085] Update the fault details table in real time and mark the newly added potential faults in the fault details table. The newly added faults refer to the power grid faults that are not involved in the training of the parent prediction model. Because with the changes in technology and equipment, new power grid faults will appear, or the fault identification characteristics of the original faults will change, resulting in new faults.

[0086] The occurrence probability and fault impact value of each new fault are counted in real time. The fault impact value is set according to the degree of impact of the new fault on the power grid when it occurs, and the value range is [0, 100]. If there is no impact, the fault impact value is 0, and the fault impact value that can cause the greatest impact on the power grid is 100. Other fault impact values ​​are set according to the impact differences to form an impact value matching table, which can be matched later or calculated in combination with interpolation. In other embodiments, the fault impact value can also be intelligently determined based on other existing technologies; for example, an impact value analysis model is established based on a neural network such as a CNN network and a DNN network, and a corresponding training set is set manually for training. The training set includes input data and output data, the input data is the fault impact data, and the output data is the fault impact value; the corresponding fault impact value is obtained by analyzing the impact value analysis model after successful training.

[0087] Allocate the newly added faults to the corresponding deviation categories according to the determination method of the deviation categories, such as allocating them to the deviation category with the highest estimated prediction accuracy;

[0088] The newly added faults in the deviation classification are marked as i, i = 1, 2, ..., n, and n is the number of newly added faults; the occurrence probability and the fault impact value are marked as λi and GZi, respectively, indicating the occurrence probability and fault impact value of the corresponding newly added faults, such as GZ1 indicating the fault impact value of the newly added fault corresponding to i = 1;

[0089] According to the formula Calculate optimization value in real time;

[0090] Where: YP is the optimized value;

[0091] When the optimization value is greater than the threshold value X1, the parent prediction model is optimized according to each newly added fault, and the parent prediction model is optimized and adjusted according to the corresponding training set set for each newly added fault. The threshold is set by technicians in this field according to the actual situation or obtained by a large amount of data simulation, that is, it is set with reference to the user's requirements for model optimization; if the user specifies several optimization situations, the corresponding optimization values ​​are calculated, and the threshold value X1 is set after each optimization value is sorted.

[0092] Step 3: According to the parent prediction model classified by deviation, a corresponding sub-prediction model is set for each prediction area, where the sub-prediction model is a power grid fault prediction model, which is the parent prediction model corresponding to the setting time; and after the parent prediction model is optimized, each sub-prediction model corresponding to the parent prediction model is updated;

[0093] Step 4: Acquire the monitoring data of the power grid area in real time, classify the monitoring data according to the regional data range of each prediction area, form the monitoring data corresponding to each sub-prediction model, mark it as regional monitoring data, analyze the regional monitoring data through the sub-prediction model, and obtain the fault prediction result. The fault prediction result includes the prediction result of each potential fault of each power grid equipment in the prediction area.

[0094] Step 5: Perform risk estimation based on the fault prediction results to obtain risk prediction results.

[0095] In one embodiment, risk estimation is performed according to the fault prediction result, and the estimation can be performed based on various existing risk estimation technologies.

[0096] In one embodiment, a method for performing risk estimation based on a fault prediction result includes:

[0097] According to the fault prediction results, the predicted potential faults are identified and marked as target faults. According to the fault prediction results corresponding to the target faults, a fault probability curve is generated, where the horizontal axis is time and the vertical axis is the fault occurrence rate, and the curve ends when the fault occurrence rate is 100%; the fault probability curve is fitted to obtain a fault probability function, and the fault probability function is marked as H(t), where t is time;

[0098] Determine the fault impact value of the target fault; calculate the risk value according to the risk assessment formula;

[0099] A risk prediction result is formed based on the risk value.

[0100] In one embodiment, the risk assessment formula is:

[0101] Wherein: FB is the risk value; GZ′ is the fault impact value; T is the interval length, that is, the time period from when the fault occurrence rate is greater than 0 to when it reaches the preset fault occurrence rate. Generally, the preset fault occurrence rate is 100%, but in some cases it may not reach 100%, so a preset value lower than 100% is set. The preset value can be marked as a predicted probability, such as 90%, 85%, etc., indicating that it is likely to occur. The shorter the interval length, the shorter the reaction time.

[0102] In one embodiment, the risk assessment formula is: FB = GZ′×max{H(t)};

[0103] Where: FB is the risk value; GZ′ is the fault impact value; max{H(t)} represents the maximum value in the fault probability function.

[0104] In one embodiment, the risk assessment formula is: FB = GZ′;

[0105] Where: FB is the risk value; GZ′ is the fault impact value.

[0106] The above formulas are all calculated by removing dimensions and taking numerical values. The formula is a formula that is closest to the actual situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained by simulating a large amount of data.

[0107] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A power grid fault prediction and risk assessment method based on deep learning, characterized in that: Methods include: Obtaining grid area information, and determining deviation classification and data dispatch range corresponding to the deviation classification according to the grid area information; A power grid fault prediction model is set for each deviation classification, which is marked as a parent prediction model; a fault detail list is obtained, and the parent prediction model is optimized in real time based on the fault detail list; the power grid area is divided according to the data dispatch range of the deviation classification to form a number of prediction areas corresponding to the deviation classification, and the data dispatch range corresponding to the prediction area is identified and marked as the regional data range; Setting a corresponding sub-prediction model for the prediction area according to the parent prediction model classified by the deviation; and updating the sub-prediction model corresponding to the parent prediction model after the parent prediction model is optimized; Acquire monitoring data of the power grid area in real time, classify the monitoring data according to the regional data range of the prediction area, form monitoring data of the prediction area, mark it as regional monitoring data, analyze the regional monitoring data through the sub-prediction model, and obtain fault prediction results; A risk prediction is performed according to the fault prediction result to obtain a risk prediction result.

2. The method for predicting and assessing power grid faults based on deep learning according to claim 1, characterized in that: Methods for partitioning power grid areas according to power grid area information include: Obtaining historical prediction records of power grid faults by each application model, wherein the application model is a power grid fault prediction model of historical application; identifying power grid faults of the power grid according to the historical prediction records, and marking the power grid faults as potential faults; Determine a deviation classification according to the historical prediction record, wherein the deviation classification is composed of potential faults, and the potential faults in different deviation classifications are different; A fault detail table is set according to the deviation classification, wherein the fault detail table is used to count potential faults and fault identification features corresponding to the potential faults; Mark corresponding potential faults for each grid device in the grid area according to the grid area information and the fault details table; determine the collection scope of the grid device marking the potential fault according to the fault identification characteristics of the potential fault; The collection range of the power grid equipment is integrated according to the deviation classification to obtain the data scheduling range corresponding to the deviation classification.

3. The method for predicting and assessing power grid faults based on deep learning according to claim 2, characterized in that: Methods for determining deviation classification based on historical forecast records include: Calculate the prediction accuracy of the application model for potential faults through historical prediction records; identify the highest prediction accuracy of the potential faults and mark it as the benchmark accuracy; calculate the standard interval of the potential faults according to the benchmark accuracy and a preset allowable deviation value; classify the potential faults according to the standard interval to obtain an initial classification corresponding to the application model, the initial classification includes potential faults whose prediction accuracy of the application model is within the standard interval; remove duplicates from the initial classification; Setting a combination standard, wherein the combination standard includes all potential faults; combining according to the combination standard and the initial classification to form a candidate combination that meets the combination standard; The candidate combinations are screened to determine a target combination; and the target combination is simulated and adjusted to obtain a deviation classification.

4. The method for predicting and assessing power grid faults based on deep learning according to claim 1, characterized in that: Methods for segmenting power grid areas according to the data dispatch range of deviation classification include: Step SA1: Divide the power grid area into several unit areas, identify the collection range of each power grid device in each unit area, and integrate them into a unit range; estimate the data analysis amount of the unit area according to the unit range; set a limit on the data amount for each deviation classification; Step SA2: Select any unit area as the initial area; when there is no initial area, end the analysis; Step SA3: Mark any unit area adjacent to the initial area as a candidate area; When there is no area to be selected, the initial area is marked as the prediction area, and the process returns to step SA2; Step SA4: Add the data analysis amount of the initial area and the data analysis amount of the selected area to obtain the amount of data to be merged, and compare the amount of data to be merged with the limited amount of data; Step SA5: when the amount of data to be merged is greater than the limited amount of data, cancel the mark of the selected area and return to step SA3; When the amount of data to be merged is not greater than the limited data amount, the initial region is merged with the selected region to form a new initial region, the amount of data to be merged is marked as the data analysis amount of the new initial region, and the process returns to step SA3.

5. The method for predicting and assessing power grid faults based on deep learning according to claim 1, characterized in that: Methods for real-time optimization of the parent prediction model based on the fault list include: The fault detail table is updated in real time, and newly added faults are identified in real time in the fault detail table; the occurrence probability and fault impact value of each newly added fault are determined in real time, and the value range of the fault impact value is [0, 100]; Assign the newly added faults to corresponding deviation categories, mark the newly added faults in the deviation categories as i, i=1, 2, ..., n, and n is the number of newly added faults; mark the occurrence probability and fault impact value as λi and GZi respectively; According to the formula Calculate the optimized value in real time; where: YP is the optimized value; Determine whether to optimize the parent prediction model corresponding to the deviation classification according to the optimization value; When it is determined that the parent prediction model is to be optimized, optimizing the parent prediction model according to the newly added fault; When it is determined that the parent prediction model is not to be optimized, no optimization operation is performed.

6. A power grid fault prediction and risk assessment method based on deep learning according to claim 5, characterized in that: Methods for risk estimation based on failure prediction results include: Identify the target fault according to the fault prediction result, and generate a fault probability curve of the target fault, wherein the horizontal axis of the fault probability curve is time and the vertical axis is the fault occurrence rate; Fitting the fault probability curve to obtain a fault probability function, marking the fault probability function as H(t), where t is time; Determine the fault impact value of the target fault; calculate the risk value according to the risk assessment formula; and form a risk prediction result based on the risk value.

7. The method for predicting and assessing power grid faults based on deep learning according to claim 6, characterized in that: The risk assessment formula is: Where: FB is the risk value; GZ′ is the fault impact value; T is the interval time.

Images